Elamur Research⌕
☰
home / papers / 2026 / Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models

Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models

Paper · 2026 · Joy Olusanya

Summary

Standard Yoruba is a low-resource language with three contrastive level tones (High, Mid, Low), written with diacritics. Without these marks, one segmental string can stand for several unrelated words. The paper asks whether keeping tone marks in transcripts helps or hurts transformer-based speech recognition. Three pre-trained systems, Meta's MMS-1B-all, OpenAI's Whisper-small and AstralZander/Yoruba_ASR, were trained and tested on two versions of the data: fully tone-marked and with tone marks removed. Two metrics were used for evaluation: TER (Tone Error Rate) alongside WER (Word Error Rate). A small word-level subset with varying degrees of tonal specification was also examined qualitatively. In every case, the unmarked condition produced clearly lower errors. With diacritics, TER rose several-fold, and outputs showed flattened tones and homophone confusions. The author links this to tokenisation, weak encoding of pitch cues and tone-agnostic pre-training. The work shows that strong multilingual ASR models do not yet handle tone. It calls for tone-aware tokenisers, alignment, training objectives and metrics for tonal low-resource languages.

Key points

  • Models compared: MMS-1B-all, Whisper-small and AstralZander/Yoruba_ASR, each evaluated on tone-marked vs. non-tone-marked Yoruba data using WER and TER.
  • TER without vs. with tone marks:
    • AstralZander: 6.44% → 30.94%
    • Whisper-small: 4.20% → 32.31%
    • MMS-1B-all: 5.20% → 52.37% (the largest increase)
  • WER grew by more than 20% for every model once diacritics were included.
  • In the word-level analysis, AstralZander preserved tone better than MMS-1B-all. MMS-1B-all often collapsed tonal contrasts, for example rendering ògùn ("medicine") as ogun, which can also mean "war" or the deity Ogun.
  • Recommendations include specialised tokenizers for tonal languages, open and well-annotated Yoruba speech corpora, and metrics beyond WER such as TER and Semantic Error Rate (SER). Stated limitations are that no dialectal variation was covered and that the pre-trained models were not optimised for tone.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksSignal Processing · 452 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Joy Olusanya. 2026. Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models. Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.loreslm-1.14

@inproceedings{olusanya2026tone,
  title = {{Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models}},
  author = {Joy Olusanya},
  booktitle = {Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loreslm-1.14},
  url = {https://doi.org/10.18653/v1/2026.loreslm-1.14}
}

References

19 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. K. Adeniyi. 2018. High tone lowering in Ìgbòmina-yorùbá. Journal of West African Languages, 45(2). find
  2. K. Adeniyi. 2021. A tonal identification of yoruba dialects. Dialectologia, 27:1–31. find
  3. A. Akinlabi. 2004. The sound system of yoruba. In N. S. Lawal, M. N. O. Sadiku, and P. A. Dopamu, editors, Understanding Yoruba Life and Culture. Africa World Press. find
  4. Alexei Baevski, Hao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33:12449–12460.
  5. W. E. Welmers. 1973. African language structures. University of California Press. find
  6. B. Connell and R. D. Ladd. 1990. Aspects of pitch realization in yoruba. Phonology, 7(1):1–29. find
  7. B. Elugbe. 1985. The tone system of ghotuo. Cambridge Papers in Phonetics and Experimental Linguistics, 4(1):1–21. find
  8. F. A. Fábùnmi. 2013. Negation in sixteen yorùbá dialects. Open Journal of Modern Linguistics, 3(1):1– 15. find
  9. F. Imam, I. Nurideen, I. Orife, and B. F. Dossou. 2025. Automatic speech recognition for african low-resource languages: Challenges and future directions. find
  10. M. Karner, S. Müller, and C. Klein. 2024. Towards improving asr outputs of spontaneous speech with llms. In Proceedings of the Annual Conference on Speech and Language Technology. find
  11. Y. O. Laniran and G. N. Clements. 2003. Downstep and high raising: Interacting factors in yoruba tone production. Journal of Phonetics, 31:203–250. find
  12. J. Levis and R. Suvorov. 2012. Automatic speech recognition. In The Encyclopedia of Applied Linguistics. Wiley. find
  13. À. Ọ. Odéjobí. 2008. Recognition of tones in Yorùbá speech: Experiments with artificial neural networks. Ph.D. thesis. find
  14. T. O. Ogunremi, K. Tubosun, A. Aremu, I. Orife, and D. I. Adelani. 2024. Ìròyìnspeech: A multi-purpose yorùbá speech corpus. In Proceedings of LREC-COLING. find
  15. Iroro F. Orife, David I. Adelani, Timi Fasubaa, Victor Williamson, Wuraola Fisayo Oyewusi, Olamilekan Wahab, and Kola Tubosun. 2020. Improving yorùbá diacritic restoration. arXiv preprint arXiv:2003.10564v1 [cs.CL]. — [arXiv:2003.10564](https://arxiv.org/abs/2003.10564) find
  16. O. Osakuade and S. King. 2024. Do discrete self-supervised representations of speech capture tone distinctions? find
  17. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2023. Whisper: General-purpose speech recognition. In Proceedings of ICML. find
  18. H. O. Raji. 2015. Ph.D. thesis. find
  19. D. R. Van Niekerk and E. Barnard. 2012. Tone realisation in a yoruba speech recognition corpus. In Proceedings of the Third International Workshop on Spoken Language Technologies for Under-resourced Languages, pages 54–59. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/olusanya-2026-tone-yoruba-asr-evaluating-impact/